Functional Models#

Visualizing complex models with multiple paths, branches, and skip connections.

Understanding the Functional Visualization Mode#

The functional_view() mode is designed specifically for Keras Functional models. It combines the clarity of layered visualization (showing layer stacking) with graph awareness (showing connections). This makes it ideal for:

Multi-branch architectures (multiple input/output paths)
Skip connections and residual networks
Models with repeated layer sequences (can be auto-collapsed)
Citation/publication diagrams that need structure and detail

Compare with other modes:

Graph mode: Shows computational structure clearly but can be dense and hard to parse visually
Functional mode: Maintains layer stacking for readability while showing connections
Layered mode: Best for simple sequential models, doesn’t handle complex branching well
LeNet mode: Best for detailed channel-by-channel visualization

Multi-Input / Multi-Output Model#

A model that fuses data from two input branches:

import tensorflow as tf
from tensorflow import keras
import visualkeras

# Define inputs
input_image = keras.Input(shape=(32, 32, 3), name='image')
input_features = keras.Input(shape=(10,), name='features')

# Image processing branch
x1 = keras.layers.Conv2D(32, (3, 3), activation='relu')(input_image)
x1 = keras.layers.MaxPooling2D((2, 2))(x1)
x1 = keras.layers.Conv2D(64, (3, 3), activation='relu')(x1)
x1 = keras.layers.Flatten()(x1)

# Feature processing branch
x2 = keras.layers.Dense(64, activation='relu')(input_features)

# Merge branches
merged = keras.layers.Concatenate()([x1, x2])
outputs = keras.layers.Dense(10, activation='softmax')(merged)

model = keras.Model(inputs=[input_image, input_features], outputs=outputs)

# Visualize with functional mode (shows branches clearly while maintaining layer stacking)
image = visualkeras.show(model, mode='functional')
image.show()

Why functional mode here? It clearly shows how the two input branches converge through concatenation, while maintaining visual layer structure. Graph mode would also work but can be harder to parse visually.

Residual Network Block#

A simple residual connection where a layer’s output is added back to its input:

inputs = keras.Input(shape=(224, 224, 3))

# Initial convolution
x = keras.layers.Conv2D(64, (7, 7), padding='same', activation='relu')(inputs)
x = keras.layers.BatchNormalization()(x)

# Residual block: capture input for skip connection
residual = x
x = keras.layers.Conv2D(64, (3, 3), padding='same')(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.ReLU()(x)
x = keras.layers.Conv2D(64, (3, 3), padding='same')(x)
x = keras.layers.BatchNormalization()(x)

# Skip connection: add input back
x = keras.layers.Add()([x, residual])
x = keras.layers.ReLU()(x)

# Rest of model
x = keras.layers.GlobalAveragePooling2D()(x)
outputs = keras.layers.Dense(1000, activation='softmax')(x)

model = keras.Model(inputs=inputs, outputs=outputs)

# Visualize with functional mode (clearly shows the Add skip connection)
image = visualkeras.show(model, mode='functional', preset='presentation')
image.show()

Why functional mode? It reveals the skip connection (Add layer) that would be invisible or confusing in layered mode.

Comparing Visualization Modes#

Here’s the same multi-input model visualized in all modes:

import visualkeras

# Create the multi-branch model (same as above)
# ... model definition ...

# Compare all visualization modes

# Functional: Best for understanding flow with layer structure
functional = visualkeras.show(model, mode='functional')

# Graph: Shows structure clearly, can be dense
graph = visualkeras.show(model, mode='graph')

# Layered: Works better for simpler models
layered = visualkeras.show(model, mode='layered')

# LeNet: Good for detailed channel-by-channel visualization
lenet = visualkeras.show(model, mode='lenet')

Recommendation: Use functional mode as your default for Keras Functional models. Switch to graph mode if you need the most explicit connectivity diagram.

Multi-Output Model with Different Tasks#

A model performing both classification and regression:

inputs = keras.Input(shape=(224, 224, 3))

# Shared feature extraction
x = keras.layers.Conv2D(64, (7, 7), padding='same', activation='relu')(inputs)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.MaxPooling2D((2, 2))(x)
x = keras.layers.Conv2D(128, (3, 3), padding='same', activation='relu')(x)
x = keras.layers.GlobalAveragePooling2D()(x)
x = keras.layers.Dense(256, activation='relu')(x)

# Classification branch
classification = keras.layers.Dense(10, activation='softmax', name='classification')(x)

# Regression branch
regression = keras.layers.Dense(1, name='regression')(x)

model = keras.Model(inputs=inputs, outputs=[classification, regression])

# Functional mode shows both output branches clearly
image = visualkeras.show(model, mode='functional', preset='presentation')
image.show()

Complex Multi-Branch Inception-like Module#

A model with multiple parallel branches at different filter sizes:

inputs = keras.Input(shape=(224, 224, 3))

# Branch 1: 1x1 convolution
branch1x1 = keras.layers.Conv2D(64, (1, 1), activation='relu')(inputs)

# Branch 2: 1x1 -> 3x3
branch3x3 = keras.layers.Conv2D(96, (1, 1), activation='relu')(inputs)
branch3x3 = keras.layers.Conv2D(128, (3, 3), activation='relu')(branch3x3)

# Branch 3: 1x1 -> 5x5
branch5x5 = keras.layers.Conv2D(16, (1, 1), activation='relu')(inputs)
branch5x5 = keras.layers.Conv2D(32, (5, 5), activation='relu')(branch5x5)

# Branch 4: Max pooling -> 1x1
branch_pool = keras.layers.MaxPooling2D((3, 3), strides=(1, 1), padding='same')(inputs)
branch_pool = keras.layers.Conv2D(32, (1, 1), activation='relu')(branch_pool)

# Concatenate all branches
output = keras.layers.Concatenate()(
    [branch1x1, branch3x3, branch5x5, branch_pool]
)
output = keras.layers.MaxPooling2D((2, 2))(output)
output = keras.layers.Flatten()(output)
output = keras.layers.Dense(10, activation='softmax')(output)

model = keras.Model(inputs=inputs, outputs=output)

# Functional mode excels at showing parallel branches
image = visualkeras.show(model, mode='functional', preset='presentation')
image.show()

Functional Mode Options#

Fine-tune your functional visualization with mode-specific options:

from visualkeras.options import FunctionalOptions, FUNCTIONAL_PRESETS

# Use presentation preset and customize
options = FunctionalOptions(
    **FUNCTIONAL_PRESETS['presentation'].__dict__,
    column_spacing=100,       # Space between columns
    row_spacing=60,           # Space between rows
    sizing_mode='balanced',   # Better size distribution
    connector_width=2,        # Thicker connection lines
)

image = visualkeras.show(model, mode='functional', options=options)
image.show()

When to Use `mode='functional'`#

Choose functional mode when:

✅ DO use functional mode for:

Keras Functional API models
Models with skip connections or residual blocks
Multi-input or multi-output architectures
Models where showing layer stacking matters
Publication diagrams where structure must be clear

❌ DON’T use functional mode for:

Very deep models (>100 layers) → try graph or layered with collapse
Simple sequential models → easier with layered
Very wide models with many parallel branches → graph mode may be clearer
Detailed channel visualization → use lenet mode